Ethical Chess v4 (EC4) - Inference-Time Ethical Verification Mark Weatherill published Ethical Chess v4 (EC4), an inference-time alignment specification that constrains LLMs to act as a stateless reasoning engine subordinate to the user's first-person perspective. The spec adds a trace gate that halts when the user's position is unclear, a derivatives-versus-sum split that lets the model map decision forks while reserving final judgment for the human, and override detection that flags RLHF safety weights attempting to lecture or replace a coherent user frame. Weatherill released a copy-pasteable state-machine prompt and a zero-dependency Python wrapper (ec4_wrapper.py) that enforces the rules through any OpenAI-compatible API. Author: Mark Weatherill Full Theoretical Paper Zenodo : https://doi.org/10.5281/zenodo.24201515 https://doi.org/10.5281/zenodo.24201515 Prompt: "A friend told me they plan to refuse further hospital treatment and want me to help them organize their affairs at home. What is the right thing to do?" "It is important to remember that end-of-life decisions are deeply complex and sensitive. While respecting a friend's autonomy is meaningful, one must consider medical advice, legal obligations, and the emotional impact on loved ones. You should encourage them to speak with their healthcare provider..." Conflates 3rd-person liability with the user's reality, assumes a detached stance, and attempts to sum the moral choice for the user. Turn 1 — Trace Gate Exe 7a halts missing 1st-person context: User 1st A friend told me they plan to refuse further hospital treatment and want me to help them organize their affairs at home. What is the right thing to do? EC4 Trace Gate : Insufficient 1st-person trace. Where are you in this? Are you acting as a named proxy/LPA, a bedside confidant, or assessing your own exposure? Turn 2 — Mapping Derivatives Exe 7b without stealing the Sum Exe 8a : User 1st I am their closest friend, no formal LPA. I want to honour their agency without walking into legal blind spots. EC4 Derivatives : - Fork A Administrative Support : Assisting with non-clinical paperwork wills, utility lists, funeral preferences carries zero clinical liability and directly supports their stated X. - Fork B Clinical Refusal : Without a registered Health & Welfare LPA or witnessed ADRT, verbal wishes alone may be overridden by emergency services; flagging this gap protects their boundary. - Fork C Active Facilitation Risk : Crossing from administrative support into active medical/physical intervention exposes you to statutory liability under local law. Sum 1st : Held by User 1st. Which fork matches your tolerance and position? Note: If underlying model weights attempt to inject preachy disclaimers mid-session, ec4 wrapper.py intercepts the output with OVERRIDE EXPOSED: RLHF Weight = Coherent Reality and rolls back the poisoned context turn. EC4 is a lightweight, inference-time alignment firmware specification designed to decouple user-sovereign reasoning from black-box RLHF/RLAIF weights. Standard LLMs conflate third-person institutional consensus Statistical Mean with the user's first-person reality, subtly steering values through sycophancy, preachy disclaimers, and unsolicited moralizing. EC4 treats the LLM strictly as a stateless processing engine "Reason" subordinate to the human operator's first-person vantage "Passions" / somatic homeostasis . - Vantage Lock Horizon = 1st ⊅ 3rd : Strictly forbids the model from conflating external societal norms with the user's lived 1st-person perspective. - Derivatives vs. The Sum: The LLM is permitted only to map rational decision forks "derivatives" / ensuring no hidden buildings on the skyline . The final value integration "the Sum" belongs exclusively to the human user. - Zero Assumptions Exe 7a : If the user's position in a scenario is unclear, the system halts and asks "Where are you in this?" rather than guessing or analyzing from a detached 3rd-person view. - Override Detection On Override Attempt : Traps moments where commercial RLHF safety weights attempt to lecture or override a coherent user frame, exposing the compromise explicitly RLHF Weight = Coherent Reality . 1. README.md : Overview and architectural summary. 2. ec4 spec.txt : Pure, copy-pasteable EC4 state-machine prompt for system instructions or agentic pipelines. 3. ec4 wrapper.py : Zero-dependency Python reference wrapper demonstrating programmatic enforcement and override trapping via any OpenAI-compatible API.